Eric W. Biederman wrote:
> Pavel Emelyanov <xemul@openvz.org> writes:
> 
>> Currently indexes for netdevices come sequentially one by
>> one, and the same stays true even for devices that are 
>> created for namespaces.
>>
>> Side effects of this are:
>>  * lo device has not 1 index in a namespace. This may break
>>    some userspace that relies on it (and AFAIR something
>>    really broke in OpenVZ VEs without this);
> 
> As it happens lo hasn't been registered first for some time
> so it hasn't had ifindex of 1 in the normal kernel.
> 
>>  * after some time namespaces will have devices with indexes
>>    like 1000000 os similar. This might be confusing for a
>>    human (tools will not mind).
> 
> Only if we wind up creating that many devices.

Nope. Create and destroy new net ns for 10000 times and you'll get it.

>> So move the (currently "global" and static) ifindex variable
>> on the struct net, making the indexes allocation look more
>> like on a standalone machine.
>>
>> Moreover - when we have indexes intersect between namespaces,
>> we may catch more BUGs in the future related to "wrong device 
>> was found for a given index".
> 
> Not yet.
> 
> I know there are several data structures internal to the kernel that
> are indexed by ifindex, and not struct net_device *.  There is the
> iflink field in struct net_device.  We need a way to refer to network
> devices in other namespaces in rtnetlink in an unambiguous way.   I
> don't see any real problems with a global ifindex assignment until
> we start migrating applications.
> 
> So please hold off on this until the kernel has been audited and
> we have removed all of the uses of ifindex that assume ifindex is
> global, that we can find.

Ok.

> Right now a namespace local ifindex seems to be just asking for
> trouble.

You said the same about caching the global pid on the task_struct,
but looks like you were wrong ;) Just kidding.

> Eric
> 
>